Metrics and Board Reporting
The coverage, efficacy, detection, response and governance metrics to report — human and non-human identity separately.
Objective
Leading indicator worth watching. Agent population growth rate versus registry coverage growth rate. When the first exceeds the second, your posture is degrading regardless of what any other metric says.
Report these separately for human and non-human identity. Refuse the combined average; it is the single most misleading number in this domain.
Coverage
- Registered agents versus discovered agents (the gap is the metric)
- Percentage of agent estate emitting action telemetry
- Percentage of agents with a named accountable human owner
- Registered agents inactive beyond their stated retirement threshold but not yet retired
- Retirements completed with a full teardown record, as a share of retirements claimed
Control efficacy
- Median and 95th-percentile credential lifetime in agent execution contexts
- Percentage of agents with zero standing credentials
- Percentage of tool calls passing through a deterministic policy engine
- Number of halts triggered by velocity or blast-radius ceilings, and false-positive rate
Detection
- Mean time to detect, from simulated agent compromise exercises
- Ratio of telemetry generated to telemetry analyzed
- Percentage of the estate covered by behavioral baselining
Response
- Kill-switch time-to-effect, by agent class, from live test
- Date of last successful test of the self-hosted forensic capability against real artifacts
- Rehearsal coverage: Tier 1 drills executed in the trailing 12 months, out of four
- Rehearsal coverage: Tier 2 tabletops executed in the trailing 12 months, out of six
- Drill pass rate on first attempt, and count of findings still open past 90 days
- Median approval review time and approval rate, from the most recent erosion probe
Governance
- Agents by autonomy tier, with trend
- Deployments blocked or tier-reduced at review
- Approval rate and median review time for human-in-the-loop controls
- Count and status of active reduced-guardrail environments
- Open exceptions by control ID and by autonomy tier, with the count past expiry
- Exceptions renewed more than once, and exceptions whose compensating control is untested